hand gesture recognition using haar-like features and a stochastic context-free grammar ieee 2008...

Hand Gesture Recognition Using Haar-Like Features and

a Stochastic Context-Free GrammarIEEE 2008

69721016 高裕凱69721043 陳思安

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

INTRODUCTION

Hand gestures are a powerful human-to-human communication modality.

Compared with traditional HCI devices (mice or keyboards), hand gestures are more convenient in exploring the 3-D virtual worlds.

The human hand is a complex articulated object consisting of many connected parts and joints.(Roughly 27 degrees of freedom)

INTRODUCTION (cont.)

Glove-based devices: Cumbersome Awkward Expensive

Early research on vision-based hand tracking usually needs the help of markers or colored gloves to make the image processing easier.

INTRODUCTION (cont.)

The research is more focused on tracking the bare hand and recognizing hand gestures without the help of any markers and gloves now.

Vision-based hand gesture recognition techniques: Appearance-based approaches 3-D hand model-based approaches

INTRODUCTION (cont.)

Appearance-based approaches: Real-time performance due to the easier 2-D

image features that are employed. 3-D hand model:

Rich description and allows a wide class of hand gestures.

Computationally expensive

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

TWO-LEVEL APPROACH

In the literature of hand gesture recognition, there are two important definitions that need to be cleared: Hand posture: Static hand pose and hand location Hand gesture: A sequence of hand postures that

are connected by continuous motions

TWO-LEVEL APPROACH (cont.)

Color-based algorithms: Distinguishing objects such as the human arm

and the face Very sensitive to lighting variations

Shape-based algorithms: Computational cost is usually too high to

implement real-time systems Requirement of noise-free image segmentation

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

POSTURE DETECTION USING HAAR-LIKE FEATURES

Originally for the task of face tracking and detection There are two motivations for the employment of the Haar-like features rather than raw pixel values. Compared with raw pixels, the Haar-like features can

efficiently reduce/increase the in-class/out-of-class variability, thus making the classification easier.

The motivation is that a Haar-like feature-based system can operate much faster than a pixelbased system.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

Each Haar-like feature consists of two or three connected “black” and “white” rectangles.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

The value of a Haar-like feature is the difference between the sums of the pixel values in the black and white rectangles,

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

The “integral image” at the location of pixel(x, y) contains the sum of the pixel values above and left of this pixel, which is inclusive,

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

Based on each Haar-like feature fj , a correspondent weak classifier hj(x) is defined by

Where x is a subwindow, and θ is a threshold. pj indicates the direction of the inequality sign.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

In practice, no single Haar-like feature can identify the object with high accuracy. However, it is not difficult to find one Haar-like feature-based classifier that has better accuracy than random guessing.

The AdaBoost learning algorithm can considerably improve the overall accuracy, stage by stage, by using a linear combination of these individually weak classifiers.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

It should be noted that a Haar-like feature could be repeatedly used in the linear combination.

The training samples are reweighted; training samples that are missed by the previous classifier are “boosted” in importance.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

To be detected by the trained cascade, the positive subwindows must pass each tage of the cascade. A negative outcome at any point leads to the immediate rejection of the subwindow

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

The reason for this strategy is based on the fact that the majority of the subwindows are negative within a single image frame, and it is a rare event for a positive subwindow to go through all of the stages.

Four postures 480 412 400 420

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

“Two fingers” positive samples

Part of the negative samples

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

A 15-stage cascade is achieved for the “two fingers” posture when the training process is terminated.

When the final required false alarm rate 1X10−6 is reached, the true-positive detection rate of the final cascade classifier is 97.5%.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

The maximum time required for the classifiers to process 100 frames was 3 s in our experiment. The classifiers had in-plane and out-of-plane rotation invariance of ±15◦.

POSTURE DETECTION USING HAAR-LIKE FEATURES (cont.)

A parallel architecture of four cascade classifiers allowed us to obtain real-time recognition of the hand postures with live inputs from theWeb camera with 15 frames/s at the resolution of 320 × 240.

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

GESTURE RECOGNITION USING AN SCFG

SCFG = stochastic(probability) context-free grammar

: finite sets of nonterminals

: finite sets of terminals

: a finite set of stochastic production rules

: start symbol,

or the combination of them)

: the probability that is associated with this production rule

GESTURE RECOGNITION USING AN SCFG (cont.)

: all of the strings that are derived from X

: a string is unambiguous and has a derivation with production rules

GESTURE RECOGNITION USING AN SCFG (cont.)

Gestures that are generated with different postures

GESTURE RECOGNITION USING AN SCFG (cont.)

GESTURE RECOGNITION USING AN SCFG (cont.)

GESTURE RECOGNITION USING AN SCFG (cont.)

GESTURE RECOGNITION USING AN SCFG (cont.)

Outline

INTRODUCTION TWO-LEVEL APPROACH POSTURE DETECTION USING HAAR-LIKE

FEATURES GESTURE RECOGNITION USING AN

SCFG CONCLUSION

CONCLUSION

In this paper, we propose a two-level approach to recognize hand gestures in real time with a singleWeb camera as the input device.

CONCLUSION (cont.)

Contributions: Real-time performance and accurate recognition

for hand postures using Haar-like features and the AdaBoost learning algorithms

Uncertain input of low-level postures, the gesture can be identified by looking for the production rule that has the highest probability

Adjusting the probability that is associated with each production rule